[
https://issues.apache.org/jira/browse/MATH-1502?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17034051#comment-17034051
]
Phil Steitz commented on MATH-1502:
-----------------------------------
I am sorry, but this is actually a regression in 3.6.1. There is a bug in the
way the new implementation handles d values that are not mass points of the KS
distribution. This should not effect actual tests, as tests will always return
attained d values. This has been fixed in Hipparchus
[https://github.com/Hipparchus-Math/hipparchus/commit/9718fe2d2842bc9dac63a355702504794454ed63]
The Hipparchus fix also improves the exact computation algorithm to be faster
and more numerically stable. I will submit a PR to make similar changes to
Commons Math if no one beats me to it.
A simpler fix would be to just copy the normalizeD method in the Hipparchus
code and recode the d values using it (leaving implementation of exactP
otherwise unchanged).
> KolmogorovSmirnovTest#exactP value is incorrect in 3.6.1 potential bug
> ----------------------------------------------------------------------
>
> Key: MATH-1502
> URL: https://issues.apache.org/jira/browse/MATH-1502
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.6.1
> Reporter: Ethan Mahintorabi
> Priority: Critical
> Attachments: exact_vs_approximate.png
>
>
> I believe that I have found a bug in the calculation of the exactP value in
> KolmogorovSmirnovTest.java in version 3.6.1.
>
> In version 3.4.1 KolmogorovSmirnovTest#exactP(0.5633587786259542, 3, 3, true)
> = 0.6
>
> Where as
>
> In version 3.6.1 KolmogorovSmirnovTest#exactP(0.5633587786259542, 3, 3, true)
> = 0.09999999999999998
>
>
> The approximateP value has stayed the same at ~=0.72. This seems to
> correspond with the change in methods of calculation in
> e38bbb9f4191d0d21dea0ba31fdc131b97a5597b
>
> [https://github.com/apache/commons-math/commit/e38bbb9f4191d0d21dea0ba31fdc131b97a5597b#diff-34100572fcc8e6537bc7445a0e82b900]
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
